Passive wireless surface acoustic wave torque sensor frequency estimation method

Abstract

The invention relates to a passive wireless surface acoustic wave torque sensor frequency estimation method, which improves the feature representation of the signal context information of a surface acoustic wave torque sensor, and enhances the understanding of a model on the dynamic change of the signal of the surface acoustic wave torque sensor. A self-attention mechanism in the deep learning network allows any position in the sequence to directly interact with all other positions, so that important information is globally balanced, and noise interference is suppressed; the distance from the root of a Root-MUSIC algorithm polynomial to a unit circle and the distance from the root to a real axis are taken as objective functions, the problem that the phase is greatly influenced by noise due to the fact that the phase is taken as the objective function traditionally is avoided, the model can adjust the root in a targeted manner, the accuracy of frequency estimation is further improved, and the estimation result is closer to a real value; according to the method, complex and expensive hardware equipment is not needed, and the construction and maintenance cost of the system is reduced; high-precision acoustic surface wave torque sensor frequency estimation has important application value in the field of rotation power of ships, automobiles and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of signal processing, in particular to a passive wireless surface acoustic wave torque sensor demodulation method. BACKGROUND

[0002] Surface acoustic wave (SAW) sensors have shown great potential in industrial torque measurement due to their passive wireless, anti-electromagnetic interference, and harsh environment resistance characteristics. The principle of torque measurement is to realize non-contact sensing of torque value by detecting the frequency shift of SAW resonator echo signal. However, the existing demodulation technology still has the following defects in complex working conditions with strong noise and weak echo signal: 1) The traditional frequency spectrum estimation algorithm has insufficient reliability in low signal-to-noise ratio (SNR) environment. The most commonly used frequency estimation method is FFT method, such as patent CN110442827B, which constructs a noise reduction matrix and then uses FFT to obtain the frequency estimation of the surface acoustic wave signal. However, FFT method is limited by the number of sampling points and has a fence effect. Root-MUSIC algorithm relies on the accurate estimation of signal covariance matrix, and locates the frequency parameter through the distribution of polynomial roots. However, in the scene of mechanical vibration, temperature fluctuation or strong electromagnetic interference, the noise-dominated covariance matrix is difficult to accurately separate the signal and noise subspaces, resulting in the divergence of polynomial roots and the significant increase of frequency estimation error. In addition, the Doppler shift generated by rotating equipment and the sudden impact load will destroy the "narrowband stationary signal" condition assumed by the algorithm, further aggravating the error accumulation, and finally making the torque measurement accuracy unable to meet the industrial demand.

[0003] 2) The emerging deep learning method tries to improve the anti-interference ability, but there is a contradiction between poor generalization and high energy consumption. Existing researches mostly use end-to-end convolutional neural network (CNN) or recurrent neural network (RNN) to directly regress frequency value, but do not fuse the physical prior knowledge of SAW device (such as temperature drift model, frequency-torque transfer function), resulting in noise sensitivity of the model and low estimation accuracy in high frequency band. At the same time, the high computational complexity of this method conflicts with the energy constraint of passive wireless sensor, causing real-time demodulation delay and making it difficult to realize practical deployment. Patent CN119535423A uses small sample learning (FSL) based on feature data set and genetic optimization algorithm to extract the optimal solution of surface acoustic wave signal feature, realizes the frequency recognition of acoustic signal feature, this method discards the traditional spectrum estimation method and relies solely on model feature learning, resulting in low feature frequency recognition accuracy, especially for weak echo signal and saturated echo signal scene, the echo feature will disappear, the recognition feature frequency error is large, and it cannot be used for high-precision surface acoustic wave sensor demodulation. SUMMARY

[0004] The present application aims at the problems of the prior art that the spectrum estimation algorithm is greatly affected by signal-to-noise ratio, poor stability and insufficient generalization ability, and proposes a passive wireless surface acoustic wave torque sensor frequency estimation method.

[0005] The technical scheme of the present application is as follows: a passive wireless surface acoustic wave torque sensor frequency estimation method, specifically comprising the following steps: 1) obtaining the echo signal of the surface acoustic wave torque sensor through the surface acoustic wave reader, and sending the echo signal to the signal input unit; the signal input unit forms a joint decision criterion by comprehensively analyzing the energy entropy and attenuation trend of the signal, frequency characteristics, and performs preliminary effectiveness judgment on the received signal to obtain effective surface acoustic wave torque sensor echo signal; 2) the effective echo signal obtained in step 1) is sent to the denoising unit of the trained deep learning CNN-GRU-Transformer network to extract the time domain characteristics of the echo signal, reconstruct the echo signal, remove the noise in the signal, and output the denoised reconstructed signal; 3) using the ROOT-MUSIC frequency estimation unit to receive the reconstructed signal to construct the signal covariance matrix, and performing eigenvalue decomposition on the signal covariance matrix to obtain the signal subspace and noise subspace, and constructing a polynomial function to obtain a polynomial root set containing the best frequency of the surface acoustic wave torque sensor signal; 4) the polynomial function obtained in step 3) is sent to the target root cuckoo optimization search and root value geometry judgment unit, and the minimum distance from the polynomial root to the unit circle and to the real axis is taken as the objective function to realize the fast search of the optimal root and complete the accurate estimation of the frequency of the surface acoustic wave torque sensor signal.

[0006] Further, the joint decision criterion: when the signal entropy energy is within the preset threshold range, the signal attenuation trend conforms to the theoretical characteristics of surface acoustic wave propagation, and the frequency is within the working bandwidth of the surface acoustic wave resonator, the signal is determined as an effective surface acoustic wave echo signal.

[0007] Further, the deep learning CNN-GRU-Transformer network denoising unit in step 2) extracts the local, time sequence and global features of the input echo signal of the surface acoustic wave torque sensor, and performs decomposition and reconstruction to obtain a noise-free surface acoustic wave torque sensor signal output, and completes the signal noise filtering.

[0008] Furthermore, the denoising unit of the deep learning CNN-GRU-Transformer network in step 2) sequentially includes a CNN sub-module unit for signal feature extraction, a GRU sub-module unit for signal temporal feature extraction, a Transformer sub-module unit for signal reconstruction and recognition, and a target function and hyperparameter optimization sub-module for fast optimization of the aforementioned network sub-module parameters. The four sub-modules are connected sequentially in the data flow to form a closed-loop structure from effective signal input, feature extraction, feature fusion, decomposition and reconstruction to model optimization.

[0009] Furthermore, in step 2), the objective function of the objective function and hyperparameter optimization submodule is to minimize the root mean square error between the echo signal of the noiseless passive wireless surface acoustic wave torque sensor and the time-domain waveform of the echo signal of the surface acoustic wave torque sensor predicted by the CNN-GRU-Transformer network.

[0010] Furthermore, in step 3), the covariance matrix calculation unit in the ROOT-MUSIC frequency estimation unit is constructed based on the discrete surface acoustic wave signal. Its eigenvalues ​​are derived from the spectral characteristics of the covariance matrix, reflecting the energy distribution of the signal in each orthogonal eigenvector direction. The eigenvalue decomposition unit uses the eigenvalue decomposition method to decompose the covariance matrix. According to the magnitude of the eigenvalues, the matrix is ​​decomposed into a signal subspace and a noise subspace, thereby realizing the distinction between the signal and noise of the surface acoustic wave torque sensor echo signal.

[0011] A passive wireless surface acoustic wave torque sensor frequency estimation system includes, in sequence, a signal input unit, a deep learning CNN-GRU-Transformer network denoising unit, a ROOT-MUSIC frequency estimation unit, a target root optimization search unit, and a root value geometric determination unit. The signal input unit receives the echo signal from the surface acoustic wave torque sensor obtained by the surface acoustic wave reader, and obtains the effective surface acoustic wave echo signal by comprehensively analyzing the energy entropy, attenuation trend and frequency characteristics of the signal. The deep learning CNN-GRU-Transformer network denoising unit is used to perform preliminary denoising on the effective surface acoustic wave echo signal and reconstruct the echo signal. The ROOT-MUSIC frequency estimation unit is used to transform the problem of searching for the spectral peaks of the reconstructed echo signal into the problem of finding the roots of a polynomial, and to obtain the set of polynomial roots containing the optimal frequency of the surface acoustic wave torque sensor signal. The target root cuckoo optimization search and root value geometry determination unit is used to perform optimal root search on the polynomial root set obtained by the ROOT-MUSIC frequency estimation unit. The search result is an accurate estimate of the frequency of the surface acoustic wave torque sensor signal. A composite objective function is constructed with the distance between the polynomial root and the unit circle and its real axis distance. The search is performed with minimizing the objective function or the root mean square form of the objective function as the search objective.

[0012] The beneficial effects of this invention are as follows: The passive wireless surface acoustic wave torque sensor frequency estimation method of this invention eliminates the need for complex and expensive hardware, reducing system construction and maintenance costs, and providing a practical technical solution for high-precision passive wireless torque sensing. Simultaneously, the high-precision surface acoustic wave torque sensor demodulation method has significant application value in fields such as temperature measurement and strain measurement. The method of this invention can meet the higher requirements for frequency estimation accuracy in these fields, providing strong support for improving the performance of related systems. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of signal transmission in the method of the present invention; Figure 2 This is a schematic diagram of the Root-MUSIC preliminary processing module of the present invention; Figure 3 This is a schematic diagram of the deep learning optimization module of the present invention; Figure 4 This is a diagram showing the internal structure of the CNN submodule, GRU submodule, Transformer submodule, objective function and hyperparameter optimization submodule based on deep learning in this invention. Detailed Implementation

[0014] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0015] This invention proposes a frequency estimation method for a passive wireless surface acoustic wave (SAW) torque sensor. Its core innovations are reflected in three aspects: First, the method combines a deep learning model (CNN-GRU-Transformer), a cuckoo-based global fast search algorithm, and the traditional Root-MUSIC frequency estimation algorithm. This fully leverages the powerful feature capture and learning capabilities of deep learning to achieve temporal feature learning and denoising of the SAW echo signal, thereby improving the signal-to-noise ratio. Second, the Root-MUSIC algorithm is used to obtain the feature root set of the SAW torque measurement signal, avoiding the problem of other spectral estimation methods being greatly affected by spectral noise during the search for the maximum value. Third, the cuckoo-based search algorithm is used with the distance from the root to the unit circle and the distance from the root to the real axis as objective functions to quickly find the optimal root, solving the problem that traditional optimization algorithms such as the bat algorithm and sparrow algorithm are prone to getting trapped in local optima. Furthermore, the phase angle of the root, which is easily affected by noise, is used as the objective function in the optimization algorithm. The above processing effectively optimizes the root estimation results of the Root-MUSIC algorithm, significantly improving the accuracy of frequency estimation for surface acoustic wave torque sensors in complex environments. Even under adverse conditions such as noise interference and multipath effects, it can more accurately estimate the frequency of surface acoustic wave torque sensor signals, enhancing the stability and reliability of the entire surface acoustic wave torque sensor demodulation system and reducing estimation errors caused by environmental changes and other factors.

[0016] like Figure 1 The schematic diagram of the signal transmission of the passive wireless surface acoustic wave torque sensor frequency estimation method of the present invention is shown. It includes a signal input unit 1, a deep learning CNN-GRU-Transformer network denoising unit 2, a ROOT-MUSIC frequency estimation unit 3, and a target root cuckoo optimization search and root value geometric determination unit 4. The signal input unit 1 receives the echo signal from the surface acoustic wave torque sensor collected by the surface acoustic wave reader and sends it to the deep learning CNN-GRU-Transformer network unit 2 for signal processing. The deep learning CNN-GRU-Transformer network unit 2 performs time-series denoising on the valid surface acoustic wave sensor echo signal obtained from the signal input unit 1 and then sends it to the ROOT-MUSIC frequency estimation unit 3. The ROOT-MUSIC frequency estimation unit 3 decomposes the denoised surface acoustic wave torque sensor signal and solves for the polynomial roots of the covariance matrix. The target root cuckoo optimization search and root value geometric determination unit 4 receives the polynomial roots solved by the ROOT-MUSIC frequency estimation unit 3 and performs a global fast optimization search, finally obtaining a high-precision frequency estimation result for the surface acoustic wave torque sensor.

[0017] The method of the present invention includes the following steps: 1. The echo signal from the surface acoustic wave (SAW) torque sensor is acquired by a SAW reader and sent to the signal input unit 1. The signal input unit 1 performs a preliminary validity assessment of the received signal to determine whether it is a valid SAW torque sensor echo signal. This assessment is based on the physical propagation characteristics of SAW and considers both signal entropy energy and signal attenuation trend. Signal entropy energy characterizes the non-uniformity of signal energy distribution, reflecting the concentration of signal energy in the time or frequency domain. A true SAW echo signal typically exhibits concentrated energy and low entropy energy; while random noise or electromagnetic interference signals have a more dispersed energy distribution and relatively high entropy energy. Simultaneously, after the excitation signal stops, the amplitude of the echo signal received by the SAW reader exhibits typical monotonic attenuation characteristics due to propagation loss; its amplitude decreases exponentially or power-lawfully with time or propagation distance. A valid echo signal should conform to this physical attenuation law, while non-target signals or noise signals often exhibit irregular fluctuations or abrupt changes. Furthermore, for sensitive surface acoustic wave devices, the operating frequency is within a certain range under certain conditions, and signals outside this range are considered interference signals.

[0018] Signal input unit 1 forms a joint judgment criterion by comprehensively analyzing the signal's energy entropy, attenuation trend, and frequency characteristics: when the signal entropy energy is within a preset threshold range (high energy concentration), the signal attenuation trend conforms to the theoretical characteristics of surface acoustic wave propagation, and the frequency is within the operating bandwidth of the surface acoustic wave resonator, the signal is determined to be a valid surface acoustic wave echo signal; otherwise, it is determined to be an invalid signal. For signals determined to be invalid, signal input unit 1 will automatically discard the data to reduce the computational load of subsequent processing units; for signals determined to be valid, they will be output to subsequent units for further signal feature extraction and analysis.

[0019] 2. When the denoising unit of the deep learning CNN-GRU-Transformer network 2 receives the valid echo signal, it extracts the temporal features of the echo signal according to the trained network structure parameters, reconstructs the signal, removes noise from the signal, and improves the signal-to-noise ratio; the denoised reconstructed signal is sent to the ROOT-MUSIC frequency estimation unit 3.

[0020] 3. The ROOT-MUSIC frequency estimation unit 3 receives the reconstructed signal and constructs the signal covariance matrix. The signal covariance matrix is ​​then decomposed into eigenvalues ​​to obtain the signal subspace and noise subspace. A polynomial function is then constructed based on the decomposed signal. The polynomial is solved to obtain the set of polynomial roots for the optimal frequency.

[0021] 4. After obtaining the polynomial root geometry, the target root cuckoo optimization search and root value geometry determination unit 4 uses the minimum distance from the polynomial root to the element circle and the distance to the real axis as the objective function to achieve a fast search for the optimal root and complete the accurate estimation of the signal frequency of the surface acoustic wave torque sensor.

[0022] The communication methods between the signal input unit 1 and the surface acoustic wave reader include, but are not limited to, RS485 communication, Wi-Fi communication, and Ethernet communication.

[0023] Depend on Figure 2 As shown, the deep learning CNN-GRU-Transformer network denoising unit 2 of the present invention is used to extract local, temporal and global features of the echo signal of the input surface acoustic wave torque sensor, and decompose and reconstruct it to obtain a noise-free surface acoustic wave torque sensor signal output, complete signal noise filtering, improve the signal-to-noise ratio, and improve the accuracy of subsequent frequency estimation algorithms, thus solving the problem of low frequency measurement accuracy caused by directly estimating the signal frequency based on neural networks.

[0024] The denoising unit 2 of the deep learning CNN-GRU-Transformer network utilizes deep learning, signal recognition, and optimization modules to construct an end-to-end multi-layer neural network structure. Internally, it sequentially includes a CNN sub-module unit 21 for signal feature extraction, a GRU sub-module unit 22 for signal temporal feature extraction, a Transformer sub-module unit 23 for signal reconstruction and recognition, and a sub-module 24 for fast optimization of the objective function and hyperparameters of the aforementioned network sub-modules. These four sub-modules are sequentially connected in the data flow, forming a closed-loop structure from effective signal input, feature extraction, feature fusion, decomposition and reconstruction to model optimization.

[0025] CNN submodule unit 21 is located at the network front end and receives the valid surface acoustic wave torque sensor echo signal confirmed by signal input unit 1. This module, based on the CNN algorithm, performs temporal local feature capture on the input surface acoustic wave torque sensor data through convolution operations. This yields information such as frequency fluctuations, abrupt changes, and attenuation of the surface acoustic wave torque sensor echo signal, completing the extraction of signal detail features within a local time window and the acquisition of multi-scale time-frequency features. Through these operations, the CNN submodule unit 21 obtains the time-frequency feature matrix. Each element reflects the signal in time. t With frequency f Energy distribution on This feature matrix contains the frequency drift rate. Amplitude attenuation coefficient and phase response Physical characteristics are used to characterize the waveform features of the echo signal from the surface acoustic wave torque sensor. The above processing completes the extraction of local, temporal, and global features from the surface acoustic wave torque sensor signal. Based on the characteristics of the surface acoustic wave signal and available computing resources, parameters such as the kernel size, number, and stride of the convolutional layers in the CNN submodule unit 21 can be dynamically adjusted to better capture features.

[0026] The GRU submodule unit 22 uses update and reset gates to obtain the short-term and long-term dependencies of the surface acoustic wave torque sensor echo signal after processing the signal by the CNN submodule unit 21. It obtains the features of the surface acoustic wave torque sensor echo information from different time dimensions through the hidden state calculation unit and outputs the results through the processed sequence feature output unit to obtain the temporal characteristics of the surface acoustic wave torque sensor echo signal. The update and reset gates in the GRU submodule unit 22 adopt an adaptive parameter adjustment strategy, which adjusts the gate weights according to the dynamic changes of the input data to improve the processing capability of sequence features. At the same time, the hyperparameters such as the number of hidden units in the GRU submodule unit 22 can be dynamically adjusted to adapt to different application scenarios.

[0027] Transformer submodule 23 fully understands the surface acoustic wave torque sensor echo signal sequence with extracted local time-frequency and temporal features through the encoder-decoder architecture, generates a context-rich output sequence, completes the processing and reconstruction of the features extracted by the aforementioned unit, and obtains a noise-free surface acoustic wave torque sensor echo signal prediction. In order to determine the magnitude of the denoising effect, it is necessary to evaluate and adjust the network parameters based on the objective function and the size of the objective function of the hyperparameter optimization submodule 24.

[0028] The objective function and hyperparameter optimization submodule 24 includes an objective function calculation unit and a genetic algorithm optimization unit. The objective function calculation unit uses the root mean square error (RMSE) of the time-domain waveform of the noiseless passive wireless surface acoustic wave torque sensor echo signal and the CNN-GRU-Transformer network-predicted surface acoustic wave torque sensor echo signal as a loss function to measure the prediction error of the deep learning model in the surface acoustic wave time-domain signal denoising process. Specifically, the RMSE calculation formula is as follows: (1), in, u 1,i,j For deep learning i One predicted surface acoustic wave time-domain waveform value u 2,i,j This represents the noise-free time-domain waveform value of the surface acoustic wave for the corresponding sample. N For the total number of samples, MLet be the time-domain waveform length of the surface acoustic wave. This is achieved by minimizing... RMSE The value can effectively improve the denoising effect of surface acoustic wave time domain waveform in complex environments, thereby achieving high-precision torque measurement.

[0029] To accelerate the model training process and obtain better deep learning network models, a genetic algorithm is introduced to encode and represent the hyperparameters of the network model (including learning rate, number of convolutional layers, number of neurons, regularization parameters, etc.). By using genetic operations such as selection, crossover, and mutation, the algorithm iteratively searches for hyperparameter combinations that minimize RMSE, thus completing the rapid optimization design of hyperparameters and solving the problem that traditional neural networks require a large amount of training to optimize hyperparameters.

[0030] The objective function and hyperparameter optimization submodule 24 also adaptively adjusts the crossover and mutation probabilities in the genetic algorithm: when the algorithm converges slowly or gets stuck in a local optimum, the system automatically increases the mutation probability to enhance search diversity; when the algorithm approaches the global optimum, it decreases the mutation probability and increases the crossover probability to accelerate convergence, thereby significantly improving the efficiency of hyperparameter optimization and speeding up the search process for the optimal hyperparameter combination.

[0031] During the training of the denoising unit 2 of the CNN-GRU-Transformer deep learning network, an early stopping strategy is adopted. When the objective function value on the validation set no longer decreases within several consecutive iterations, training is stopped to prevent overfitting and improve the network's generalization ability.

[0032] The denoising unit 2 of the deep learning CNN-GRU-Transformer network adopts a hybrid precision training strategy, which reduces memory usage and training time while ensuring model training accuracy. When the amount of surface acoustic wave torque sensor signal data is too large and the processing capacity of a single deep learning CNN-GRU-Transformer network is insufficient, a distributed computing architecture can be adopted to deploy multiple deep learning CNN-GRU-Transformer networks in parallel to improve processing efficiency.

[0033] Depend on Figure 3 As shown, the ROOT-MUSIC frequency estimation unit 3 of the present invention constructs a signal covariance matrix, obtains the signal subspace and noise subspace through eigenvalue decomposition, and constructs a polynomial function to obtain a polynomial root set containing the optimal frequency of the surface acoustic wave torque sensor signal. The ROOT-MUSIC frequency estimation unit 3 includes a covariance matrix calculation unit 31, an eigenvalue decomposition unit 32, a polynomial function construction unit 33, and a polynomial root set output unit 34.

[0034] The covariance matrix calculation unit 31 processes the input discretized surface acoustic wave signal and calculates the covariance matrix using formula (2). (2), in The first of the surface acoustic wave torque sensor n There are discrete sampling points, which can be real or complex numbers; N is the total number of sampling points; the signal length is... M .

[0035] The variance matrix C is obtained: (3), The eigenvalue decomposition unit 32 decomposes the covariance matrix C using the eigenvalue decomposition (EVD) method. This matrix is ​​constructed by the covariance matrix calculation unit 31 based on the discrete surface acoustic wave signal, therefore its eigenvalues... The spectral characteristics derived from this covariance matrix reflect the energy distribution of the signal along the directions of each orthogonal eigenvector. This is achieved through calculation... The eigenvector matrix is ​​obtained. With the corresponding eigenvalue diagonal matrix These eigenvalues ​​represent the second-order statistical energy characteristics of the surface acoustic wave (SAW) signal in different eigenvector directions. Larger eigenvalues ​​correspond to a signal subspace with significant SAW torque sensor echo signals, while smaller and approximately equal eigenvalues ​​correspond to a noise subspace in the SAW torque sensor echo signals. Therefore, the matrix can be decomposed into signal subspaces based on the magnitude of the eigenvalues. With noise subspace This enables the differentiation between the echo signal and noise from the surface acoustic wave torque sensor.

[0036] The polynomial function of the Root-MUSIC algorithm is constructed in polynomial function construction unit 33, and solved using appropriate numerical solution methods (such as improved numerical iterative algorithms) to obtain the roots of the polynomial. Specifically, the root of the polynomial is obtained by using the signal subspace projection matrix. Based on this, and combined with the array steering vector (4), construct a structure about the complex variables. Polynomial function (5): (4), (5).

[0037] The coefficients in this polynomial are determined by The matrix elements are determined, representing the projected energy of the signal subspace in different frequency directions. The complex roots of this polynomial can be obtained by solving it using adjoint matrix eigenvalue decomposition or an improved numerical iterative algorithm. Among these roots, the root with a magnitude closest to 1 and closest to the real axis corresponds to the true frequency component of the signal, and its phase angle is... Each polynomial root set corresponds one-to-one with the resonant frequency of surface acoustic waves. The polynomial root set output unit 34 normalizes the polynomial roots, ensuring the data is on a uniform scale and improving the processing efficiency and frequency estimation accuracy of each submodule. The input to each unit is the set of complex roots obtained by the polynomial function construction unit 33. This set contains mixed information of signal frequency components and noise components. The unit first filters elements closest to the unit circle based on the magnitude of the root. ) effective signal root To eliminate spurious roots caused by noise, the selected roots are then normalized using the unit circle normalization method, and their amplitudes are calculated. (6) This ensures that all valid roots are projected onto the unit circle, thereby achieving a unified scale representation of frequency characteristics under different signal samples and operating conditions. The normalized root set Only the phase information is retained, including its phase angle. Each frequency corresponds one-to-one with the resonant frequency of surface acoustic waves. The output of this unit is a normalized frequency characteristic set. The calculation formula is as follows: (7), of which This represents the system sampling frequency. The output signal is in the form of a frequency feature vector, which can be directly input into subsequent frequency estimation and feature recognition modules for modeling and optimization. Through the normalization processing of this unit, not only is the computational stability and training efficiency of each submodule improved at a uniform scale, but the robustness and accuracy of surface acoustic wave frequency estimation are also effectively enhanced.

[0038] Depend on Figure 4 As shown, the target root cuckoo optimization search and root value geometric determination unit 4 of the present invention is used to output the complex root set of the polynomial root set output unit 34. In this algorithm, the optimal signal root is determined quickly based on geometric criteria. The unit uses the unit circle and real axis on the complex plane in the Root-MUSIC algorithm as geometric references, where the unit circle originates from the aforementioned polynomial root normalization process (representing all normalized complex roots). All meet (corresponding to the actual frequency distribution trajectory of the signal), while the real axis serves as the phase reference axis in the complex plane. This reflects the symmetry characteristics of the phase of the main frequency component of the signal.

[0039] The target root cuckoo optimization search and root value geometric determination unit 4 includes a cuckoo optimization search unit 41 and a root value geometric determination unit 42. The cuckoo optimization search unit 41 is based on the improved cuckoo search algorithm (CSA), constructing a composite objective function using the distance between the polynomial root and the unit circle and its real axis distance. (8), The first term represents the degree of deviation of the root from the unit circle, and the second term represents the perpendicular distance of the root relative to the real axis. , These are the weighting coefficients. The algorithm aims to minimize the objective function. or its root mean square form To find the target, a Levy flight and an adaptive step size mechanism are used to iteratively optimize in the complex plane, quickly locating the optimal polynomial root that minimizes both distances simultaneously.

[0040] In the process of searching for the best polynomial root in the Cuckoo Optimization Search Unit 41, the adjustment rules for the search step size are improved, and a minimum step size adjustment threshold is set to avoid the problem of excessively long search process.

[0041] Preferably, in the rooted geometry determination unit 42, the root mean square of the distance between the polynomial root and the unit circle and the distance between the polynomial root and the real axis is used as the objective function. This addresses the one-sidedness of using only the unit distance as a criterion in the traditional root-finding process, and also solves the problem of poor results caused by the phase angle being greatly affected by noise when using the phase angle as the objective function. The optimized root is output for subsequent frequency calculation, thereby improving the accuracy of frequency estimation of the surface acoustic wave torque sensor.

[0042] The frequency estimation method for surface acoustic wave torque sensor of the present invention: (1) Under various signal-to-noise ratio conditions, the present invention can better preserve the useful components of the signal and remove noise, showing higher peak signal-to-noise ratio and structural similarity; (2) Improve the feature representation of the signal context information of surface acoustic wave torque sensor, and enhance the model's understanding of the dynamic changes of surface acoustic wave torque sensor signal; (3) The self-attention mechanism in the deep learning network allows any position in the sequence to interact directly with all other positions, thereby globally weighing important information and suppressing noise interference; (4) The root-MUSIC algorithm is used instead of the traditional MUSIC algorithm for frequency estimation, transforming the peak search problem into the problem of finding the root of a polynomial, avoiding the search process and improving computational efficiency; (5) The Cuckoo algorithm with global fast search capability is used to automatically find the optimal frequency root, solving the current problem. (6) Using intelligent optimization algorithms such as genetic algorithms and particle swarm optimization to optimize the root search process of root-MUSIC, problems such as premature convergence, getting trapped in local optima, and parameter sensitivity may exist; (7) Using the distance from the root of the polynomial of the Root-MUSIC algorithm to the unit circle and the distance from the root to the real axis as the objective function, the problem of phase being greatly affected by noise caused by the traditional objective function of phase is avoided. The model can adjust the root in a targeted manner, further improving the accuracy of frequency estimation and making the estimation result closer to the true value; (8) The method of the present invention does not require complex and expensive hardware equipment, reducing the construction and maintenance costs of the system; High-precision surface acoustic wave torque sensor frequency estimation has important application value in the field of rotating power such as ships and automobiles. The method of the present invention can meet the higher requirements of frequency estimation accuracy in these fields and provide strong support for the performance improvement of related systems.

[0043] The embodiments described above merely illustrate specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A frequency estimation method for a passive wireless surface acoustic wave torque sensor, characterized in that, Specifically, the steps are as follows: 1) Acquire the echo signal from the surface acoustic wave (SAW) torque sensor using a SAW reader, and send the echo signal to the signal input unit. The signal input unit comprehensively analyzes the energy entropy, attenuation trend, and frequency characteristics of the signal to form a joint judgment criterion, and performs a preliminary validity judgment on the received signal to obtain a valid SAW torque sensor echo signal; 2) The valid echo signal obtained in step 1) is sent to the denoising unit of the trained deep learning CNN-GRU-Transformer network to extract the time-domain features of the echo signal, reconstruct the echo signal, and remove noise from the signal. 3) Use the ROOT-MUSIC frequency estimation unit to receive the reconstructed signal, construct the signal covariance matrix, and perform eigenvalue decomposition on the signal covariance matrix to obtain the signal subspace and noise subspace. Construct a polynomial function to obtain the set of polynomial roots corresponding to the optimal frequency of the surface acoustic wave torque sensor signal. 4) Send the polynomial function obtained in step 3) into the target root cuckoo optimization search and root value geometry determination unit. Use the minimum distance from the polynomial root to the unit circle and the minimum distance to the real axis as the objective function to achieve a fast search for the optimal root and complete the accurate estimation of the frequency of the surface acoustic wave torque sensor signal.

2. The frequency estimation method for a passive wireless surface acoustic wave torque sensor according to claim 1, characterized in that, The joint judgment criterion is as follows: when the signal entropy energy is within a preset threshold range, the signal attenuation trend conforms to the theoretical characteristics of surface acoustic wave propagation, and the frequency is within the operating bandwidth of the surface acoustic wave resonator, the signal is determined to be a valid surface acoustic wave echo signal.

3. The frequency estimation method for a passive wireless surface acoustic wave torque sensor according to claim 1 or 2, characterized in that, In step 2), the deep learning CNN-GRU-Transformer network denoising unit extracts local, temporal, and global features of the echo signal from the input surface acoustic wave torque sensor, and decomposes and reconstructs them to obtain a noise-free surface acoustic wave torque sensor signal output, thus completing the signal noise filtering.

4. The frequency estimation method for a passive wireless surface acoustic wave torque sensor according to claim 3, characterized in that, The denoising unit of the deep learning CNN-GRU-Transformer network in step 2) includes a CNN sub-module unit for signal feature extraction, a GRU sub-module unit for signal temporal feature extraction, a Transformer sub-module unit for signal reconstruction and recognition, and a target function and hyperparameter optimization sub-module for fast optimization of the parameters of the aforementioned network sub-modules. The four sub-modules are connected in sequence on the data flow to form a closed-loop structure from effective signal input, feature extraction, feature fusion, decomposition and reconstruction to model optimization.

5. The frequency estimation method for a passive wireless surface acoustic wave torque sensor according to claim 4, characterized in that, In step 2), the objective function and hyperparameter optimization submodule aims to minimize the root mean square error between the echo signal of the noiseless passive wireless surface acoustic wave torque sensor and the time-domain waveform of the echo signal predicted by the CNN-GRU-Transformer network.

6. The frequency estimation method for a passive wireless surface acoustic wave torque sensor according to claim 1, characterized in that, In step 3), the covariance matrix calculation unit in the ROOT-MUSIC frequency estimation unit is constructed based on the discrete surface acoustic wave signal. Its eigenvalues ​​are derived from the spectral characteristics of the covariance matrix, reflecting the energy distribution of the signal in each orthogonal eigenvector direction. The eigenvalue decomposition unit uses the eigenvalue decomposition method to decompose the covariance matrix. According to the magnitude of the eigenvalues, the matrix is ​​decomposed into a signal subspace and a noise subspace, thereby realizing the distinction between the signal and noise of the surface acoustic wave torque sensor echo signal.

7. A passive wireless surface acoustic wave torque sensor frequency estimation system, characterized in that, The system sequentially includes a signal input unit, a deep learning CNN-GRU-Transformer network denoising unit, a ROOT-MUSIC frequency estimation unit, a target root cuckoo optimization search unit, and a root value geometric determination unit. The signal input unit receives the echo signal from the surface acoustic wave torque sensor obtained through a surface acoustic wave reader, and obtains the effective surface acoustic wave echo signal by comprehensively analyzing the signal's energy entropy, attenuation trend, and frequency characteristics. The deep learning CNN-GRU-Transformer network denoising unit is used to perform preliminary denoising on the effective surface acoustic wave echo signal and reconstruct the echo signal. The ROOT-MUSIC frequency estimation unit transforms the spectral peak search problem of reconstructing the echo signal into the problem of finding the roots of a polynomial, obtaining a set of polynomial roots corresponding to the optimal frequency of the surface acoustic wave torque sensor signal. The target root cuckoo optimization search and root value geometric determination unit performs optimal root search on the set of polynomial roots obtained by the ROOT-MUSIC frequency estimation unit, and the search result is an accurate estimate of the frequency of the surface acoustic wave torque sensor signal. A composite objective function is constructed using the distance between the polynomial roots and the unit circle and its real axis distance, and the search is performed with minimizing the objective function or the root mean square form of the objective function as the search objective.

Images